Group: Stanley Wei & Ryu Adams
For our final project for UCLA's COM SCI 188: Introduction to Robotics course, we implemented two methods for imitation learning from expert demonstrations:
- DMP: Cartesian-space dynamic motion primitives.
- NN: Behavior cloning using a neural network.
Evaluated on the robosuite Nut Assembly task:
| Method | Training Time | Success Rate | Avg. Time |
|---|---|---|---|
| DMP | 378 ms. (1 demo.) | 0.76 | 6.389 sec. |
| NN | 3.2 min (200 demos) | 0.72 | 3.481 sec. |
You can see our project website here: https://sites.google.com/g.ucla.edu/stanley-ryu-cs188-project
This project requires robosuite to be installed. Additionally, torch is required to run the neural network components.
To test the DMP policy, run:
cd dmp && python3 test_dmp.py
- Note: The
DMPPolicyclass requires a file containing demonstrations to learn from!
To test the NN policy, run:
cd nn && python3 test_nn.py
- To train your own NN policy, use
behavior_cloning.py!